Training Generative adversarial networks (GANs) is a challenging task. The generator in GANs transform noise, typically Gaussian distributed, into realistic images. In this work, we proposed training GANs with images as inputs, but without enforcing any pairwise constraints. The process is made efficient by identifying closely related “friendly neighborhood” datasets of the target, inspiring the moniker, Spider GAN. The friendly neighbors are identified via a newly proposed polyharmonic-spline kernel-based measure called the signed inception distance (SID). Spider GANs results in faster convergence, as the generator can discover correspondence even between seemingly unrelated datasets, for instance, between objects and faces. Cascading Spider variants of StyleGAN3 with pre-trained GAN generators, a new flavor of transfer learning, achieves state-of-the-art Fréchet inception distance (FID) values in one-fifth of the training iterations compared to the baselines, on high-resolution small datasets such as MetFaces, Ukiyo-E Faces and AFHQ.
S. Asokan and C. S. Seelamantula, “Spider GAN: Leveraging Friendly Neighbors to Accelerate GAN Training,” In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2023, Vancouver, Canada